SCAIL-2 GGUF

Rebels GGUF quantizations of SCAIL-2, the end-to-end character-animation / video motion-transfer model (Wan 2.1 14B backbone) from zai-org. These run the SCAIL-2 DiT in ComfyUI at a fraction of the VRAM the full fp16/fp8 weights require.

Quantized by RealRebelAI · GitHub · YouTube

⚡ Load with the GGUF Unet Loader (city96's ComfyUI-GGUF — Unet Loader (GGUF)). Place the .gguf in ComfyUI/models/unet/.

Quant tiers

• Q2_K 6 GB - Smallest, runs on minimal VRAM, expect quality loss

• Q3_K_M 8 GB - Budget tier, better coherence than Q2

• Q4_K_M 10 GB - Recommended daily driver

• Q5_K_M 12 GB - Sweet spot above Q4

• Q6_K 14 GB - Higher fidelity

• Q8_0 17 GB - Closest to fp16

The loader memory-maps the model, so a larger file costs disk and streaming time, not resident RAM.

Required files

Download each of these separately and place them in the listed ComfyUI folder.

😎 Model

ComfyUI/models/unet/

https://huggingface.co/realrebelai/SCAIL-2_GGUF/tree/main

📝 Text Encoder

ComfyUI/models/text_encoders/ https://huggingface.co/chatpig/encoder/blob/main/umt5_xxl_fp8_e4m3fn_scaled.safetensors

🎛️ LoRA (LightX2V step/cfg distill)

ComfyUI/models/loras/ https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v/blob/main/loras/Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors

🎯 SAM 3.1 Multiplex

ComfyUI/models/sam/ https://huggingface.co/Comfy-Org/sam3.1/blob/main/checkpoints/sam3.1_multiplex_fp16.safetensors

👁️ CLIP Vision

ComfyUI/models/clip_vision/ https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/clip_vision/clip_vision_h.safetensors

🎨 VAE

ComfyUI/models/vae/ https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors

➕ Optional: SCAIL-2 DPO LoRA (untested)

ComfyUI/models/loras/ https://huggingface.co/Comfy-Org/SCAIL-2/blob/main/loras/wan2.1_SCAIL_2_DPO_lora_bf16.safetensors

Folder structure

ComfyUI/models/
├── unet/
│   └── SCAIL-2-Q4_K_M.gguf            ← from this repo
├── text_encoders/
│   └── umt5-xxl-enc-fp8_e4m3fn.safetensors
├── loras/
│   ├── Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64.safetensors
│   └── wan2.1_SCAIL_2_DPO_lora_bf16.safetensors   (optional)
├── sam/
│   └── sam3.1_multiplex_fp16.safetensors
├── clip_vision/
│   └── clip_vision_h.safetensors
└── vae/
    └── wan_2.1_vae.safetensors

Notes

• WEIGHT NOT MERGED warning on patch_embedding is harmless. ComfyUI builds a 36-channel patch embedding and concatenates the mask channels at runtime; the model fills them internally. The stored 20-channel weight is expected. Generation proceeds normally.

• The colored mask is a required input even in single-character Animation Mode — don't remove it from the workflow.

• Set width and height explicitly (both divisible by 16; 832×480 is a good 480p start).

• The SCAIL2ColoredMask node may require a recent / nightly ComfyUI build.

Credits

• Model: zai-org / SCAIL-2

• GGUF quantization: RealRebelAI

• GGUF tooling: city96 / ComfyUI-GGUF